InstCombine: Preserve nuw when reassociating nuw ops [3/3]

[arsenm]

Summary

Alive says this is OK.

[lebedev.ri]

Can this end up producing anything other than mul-of-add that we care about?
https://rise4fun.com/Alive/WtnX

lib/Transforms/InstCombine/InstructionCombining.cpp

595

Outdated comment?

620–621

Why is this restricted to m_APInt()?
https://rise4fun.com/Alive/WtnX

[lebedev.ri]

(just marking as reviewed)

[arsenm]

Don't restrict to just constant

[lebedev.ri]

Looks good.

lib/Transforms/InstCombine/InstructionCombining.cpp

595

Still outdated?

[arsenm]

r364235

lib/Transforms/InstCombine/InstructionCombining.cpp

if (B == D || (InnerCommutative && B == C)) {
       }
     }
 
   if (SimplifiedInst) {
     ++NumFactor;
     SimplifiedInst->takeName(&I);
 
     // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
     // TODO: Check for NUW.
     if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
       if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
         bool HasNSW = false;
-        if (isa<OverflowingBinaryOperator>(&I))
+        bool HasNUW = false;
+        if (isa<OverflowingBinaryOperator>(&I)) {
           HasNSW = I.hasNoSignedWrap();
+          HasNUW = I.hasNoUnsignedWrap();
+        }
 
-        if (auto *LOBO = dyn_cast<OverflowingBinaryOperator>(LHS))
+        if (auto *LOBO = dyn_cast<OverflowingBinaryOperator>(LHS)) {
           HasNSW &= LOBO->hasNoSignedWrap();
+          HasNUW &= LOBO->hasNoUnsignedWrap();
+        }
 
-        if (auto *ROBO = dyn_cast<OverflowingBinaryOperator>(RHS))
+        if (auto *ROBO = dyn_cast<OverflowingBinaryOperator>(RHS)) {
           HasNSW &= ROBO->hasNoSignedWrap();
+          HasNUW &= ROBO->hasNoUnsignedWrap();
+        }
 
+        const APInt *CInt;
+        if (TopLevelOpcode == Instruction::Add &&
+            InnerOpcode == Instruction::Mul) {
           // We can propagate 'nsw' if we know that
           //  %Y = mul nsw i16 %X, C
           //  %Z = add nsw i16 %Y, %X
           // =>
           //  %Z = mul nsw i16 %X, C+1
           //
           // iff C+1 isn't INT_MIN
-        const APInt *CInt;
-        if (TopLevelOpcode == Instruction::Add &&
-            InnerOpcode == Instruction::Mul)
-          if (match(V, m_APInt(CInt)) && !CInt->isMinSignedValue())
+          if (match(V, m_APInt(CInt))) {
+            if (!CInt->isMinSignedValue())
               BO->setHasNoSignedWrap(HasNSW);
           }
+
+          // nuw can be propagated with any constant or nuw value.
+          BO->setHasNoUnsignedWrap(HasNUW);
+        }
+      }
     }
   }
   return SimplifiedInst;
 }
 
 /// This tries to simplify binary operations which some other binary operation
 /// distributes over either by factorizing out common terms
 /// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in

test/Transforms/InstCombine/reassociate-nuw.ll

;
   %sub0 = sub nuw i32 %x, 4
   %sub1 = sub nuw i32 %y, 8
   %sub2 = sub nuw i32 %sub0, %sub1
   ret i32 %sub2
 }
 
 define i32 @tryFactorization_add_nuw_mul_nuw(i32 %x) {
 ; CHECK-LABEL: @tryFactorization_add_nuw_mul_nuw(
-; CHECK-NEXT:    [[ADD2:%.*]] = shl i32 [[X:%.*]], 2
+; CHECK-NEXT:    [[ADD2:%.*]] = shl nuw i32 [[X:%.*]], 2
 ; CHECK-NEXT:    ret i32 [[ADD2]]
 ;
   %mul1 = mul nuw i32 %x, 3
   %add2 = add nuw i32 %mul1, %x
   ret i32 %add2
 }
 
 define i32 @tryFactorization_add_nuw_mul_nuw_int_max(i32 %x) {
 ; CHECK-LABEL: @tryFactorization_add_nuw_mul_nuw_int_max(
-; CHECK-NEXT:    [[ADD2:%.*]] = shl i32 [[X:%.*]], 31
+; CHECK-NEXT:    [[ADD2:%.*]] = shl nuw i32 [[X:%.*]], 31
 ; CHECK-NEXT:    ret i32 [[ADD2]]
 ;
   %mul1 = mul nuw i32 %x, 2147483647
   %add2 = add nuw i32 %mul1, %x
   ret i32 %add2
 }
 
 define i32 @tryFactorization_add_mul_nuw(i32 %x) {

 ; CHECK-LABEL: @tryFactorization_add_nuw_mul(
 ; CHECK-NEXT:    [[ADD2:%.*]] = shl i32 [[X:%.*]], 2
 ; CHECK-NEXT:    ret i32 [[ADD2]]
 ;
   %mul1 = mul nuw i32 %x, 3
   %add2 = add i32 %mul1, %x
   ret i32 %add2
 }
+
+define i32 @tryFactorization_add_nuw_mul_nuw_mul_nuw_var(i32 %x, i32 %y, i32 %z) {
+; CHECK-LABEL: @tryFactorization_add_nuw_mul_nuw_mul_nuw_var(
+; CHECK-NEXT:    [[MUL21:%.*]] = add i32 [[Y:%.*]], [[Z:%.*]]
+; CHECK-NEXT:    [[ADD1:%.*]] = mul nuw i32 [[MUL21]], [[X:%.*]]
+; CHECK-NEXT:    ret i32 [[ADD1]]
+;
+  %mul1 = mul nuw i32 %x, %y
+  %mul2 = mul nuw i32 %x, %z
+  %add1 = add nuw i32 %mul1, %mul2
+  ret i32 %add1
+}
+
+define i32 @tryFactorization_add_nuw_mul_mul_nuw_var(i32 %x, i32 %y, i32 %z) {
+; CHECK-LABEL: @tryFactorization_add_nuw_mul_mul_nuw_var(
+; CHECK-NEXT:    [[MUL21:%.*]] = add i32 [[Y:%.*]], [[Z:%.*]]
+; CHECK-NEXT:    [[ADD1:%.*]] = mul i32 [[MUL21]], [[X:%.*]]
+; CHECK-NEXT:    ret i32 [[ADD1]]
+;
+  %mul1 = mul i32 %x, %y
+  %mul2 = mul nuw i32 %x, %z
+  %add1 = add nuw i32 %mul1, %mul2
+  ret i32 %add1
+}
+
+define i32 @tryFactorization_add_nuw_mul_nuw_mul_var(i32 %x, i32 %y, i32 %z) {
+; CHECK-LABEL: @tryFactorization_add_nuw_mul_nuw_mul_var(
+; CHECK-NEXT:    [[MUL21:%.*]] = add i32 [[Y:%.*]], [[Z:%.*]]
+; CHECK-NEXT:    [[ADD1:%.*]] = mul i32 [[MUL21]], [[X:%.*]]
+; CHECK-NEXT:    ret i32 [[ADD1]]
+;
+  %mul1 = mul nuw i32 %x, %y
+  %mul2 = mul i32 %x, %z
+  %add1 = add nuw i32 %mul1, %mul2
+  ret i32 %add1
+}
+
+define i32 @tryFactorization_add_mul_nuw_mul_var(i32 %x, i32 %y, i32 %z) {
+; CHECK-LABEL: @tryFactorization_add_mul_nuw_mul_var(
+; CHECK-NEXT:    [[MUL21:%.*]] = add i32 [[Y:%.*]], [[Z:%.*]]
+; CHECK-NEXT:    [[ADD1:%.*]] = mul i32 [[MUL21]], [[X:%.*]]
+; CHECK-NEXT:    ret i32 [[ADD1]]
+;
+  %mul1 = mul nuw i32 %x, %y
+  %mul2 = mul nuw i32 %x, %z
+  %add1 = add i32 %mul1, %mul2
+  ret i32 %add1
+}